Interpretable stance detection in social media via topic-guided transformers

Stance detection in social media is a critical task in computational social science, supporting the analysis of ideological polarization, public opinion, and sociopolitical discourse. However, the brevity, noise, and contextual ambiguity of user-generated content pose significant challenges for traditional NLP models, while large language models (LLMs), despite their impressive zero-shot capabilities, often act as opaque black boxes with limited interpretability and domain adaptability. To overcome these limitations, we propose an interpretable hybrid framework that combines BERTopic for unsupervised semantic topic discovery with RoBERTa for sentiment-informed stance classification. Our approach explicitly fuses latent topical structure, sentiment polarity, and deep contextual embeddings, enabling both high predictive accuracy and transparent explanatory insights. Extensive evaluations on SemEval-2016 Task 6 and the COVID-19 Stance Dataset demonstrate that our model achieves macro-F1 scores of 78.4% and 77.2%, surpassing competitive baselines such as TextCNN, BiLSTM-Attention, fine-tuned BERT, and CT-BERT. Topic coherence metrics (NPMI, UCI, UMass) further confirm BERTopic’s superiority over Structural Topic Modeling (STM), underscoring its effectiveness in short-text settings. Beyond quantitative results, we enhance interpretability through topic–stance heatmaps, qualitative case studies, and a human-centered evaluation, while ablation analysis validates the contribution of each pipeline component. Finally, we discuss ethical and societal risks, cost considerations, and real-world deployment implications, and outline future directions including multilingual expansion, real-time stance monitoring, and human-in-the-loop integration. This work advances explainable social AI by bridging performance and interpretability in stance detection, making it a powerful tool for opinion mining, policy analysis, and misinformation tracking.